FinDialogLens preprint reports fewer AI calls for financial chat analysis

Researchers presenting FinDialogLens report that routing financial quote requests between a rule-based engine and a language model cut model calls for final-price extraction by 85%, saving over $300/day at their 70,000-RFQ/day scale. The preprint addresses chatrooms where prices and trade outcomes appear across interleaved conversations. The pipeline uses compact classifiers to identify requests for quotes and relevant messages, groups messages by request, and extracts trade details. With GPT-4o, it achieved 92.1% accuracy on final prices and 94.3% on trade outcomes. The cheaper routing approach recovered half of the accuracy gap between the rule-based engine and the full GPT-4o pipeline.